Paper Summary: Recent AI Advances in Financial Time-Series (Sep 6‑12, 2025)

Paper 1: "Text‑Trained LLMs Can Zero‑Shot Extrapolate PDE Dynamics" (arXiv: http://arxiv.org/pdf/2509.06322v1) – Authors: Jiajun Bao, Nicolas Boullé, Toni J. B. Liu, Raphaël Sarfati, Christopher J. Earls. The authors show that a text‑trained large language model can infer spatio‑temporal dynamics from discretized PDE solutions without any fine‑tuning or natural‑language prompting. Prediction accuracy improves as the temporal context lengthens but degrades with finer spatial discretization. In multi‑step forecasting the model recursively predicts future spatial states, and the error grows algebraically with the prediction horizon, mirroring the global error accumulation of classic finite‑difference solvers. They attribute these trends to a “contextual neural extrapolation law” that links prediction quality to context length and output length, and they analyze token‑level output distributions to understand how the LLM internally processes PDE solutions.

Paper 2: "time2time: Causal Intervention in Hidden States to Simulate Rare Events in Time Series Foundation Models" (arXiv: http://arxiv.org/pdf/2509.05801v1) – Authors: Debdeep Sanyal, Aaryan Nagpal, Dhruv Kumar, Murari Mandal, Saurabh Deshpande. Transformer‑based foundation models excel at regular pattern prediction, yet it is unclear whether they internalize semantic concepts such as market regimes or merely fit curves, and whether their hidden states can be used to simulate rare, high‑risk events. The authors introduce **activation transplant**, a causal intervention that imposes the statistical moments of a target event (e.g., a historical market crash) onto the hidden states during the forward pass. Injecting crash semantics steers the model to forecast recessions, while injecting calm semantics suppresses crash predictions and restores stability. Experiments reveal that the models encode a graded notion of event severity: the norm of the latent vector correlates with the magnitude of systemic shocks. Validation on two structurally different time‑series foundation models—Toto (decoder‑only) and Chronos (encoder‑decoder)—demonstrates that manipulable, semantically grounded representations are a robust property of large‑scale time‑series transformers, enabling direct “if‑analysis” for strategic stress testing.

Paper 3: "Of Graphs and Tables: Zero‑Shot Node Classification with Tabular Foundation Models" (arXiv: http://arxiv.org/pdf/2509.07143v1) – Authors: Adrian Hayler, Xingyue Huang, İsmail İlkan Ceylan, Michael Bronstein, Ben Finkelshtein. Graph foundation models (GFMs) often train on synthetic graph datasets that do not reflect real‑world graph distributions, limiting their generalization. By contrast, tabular foundation models (TFMs) excel on traditional tabular prediction tasks and transfer well to domains such as time‑series, NLP, and vision. The authors recast node classification as a tabular problem: each node becomes a row whose columns contain feature, structural, and label information. This allows TFMs to perform zero‑shot node classification via contextual learning. They propose **TabGFM**, a framework that first converts a graph into a table using feature and structure encoders, then applies multiple TFMs to different sub‑sampled tables, and finally aggregates the outputs. Experiments on 28 real‑world datasets show that TabGFM consistently improves over task‑specific GNNs and state‑of‑the‑art GFMs, highlighting the potential of table reconstruction for scalable and generalizable graph learning.

Paper 4: "FinZero: Launching Multi‑modal Financial Time Series Forecast with Large Reasoning Model" (arXiv: http://arxiv.org/pdf/2509.08742v1) – Authors: Yanlong Wang, Jian Xu, Fei Ma, Hongkang Zhang, Hang Yu, Tiantian Gao, Yu Wang, Haochen You, Shao‑Lun Huang, Danny Dongning Sun, Xiao‑Ping Zhang. Financial time‑series forecasting is both crucial and challenging. Traditional pipelines standardize the series before modeling, which discards important information, and many models require fixed numbers of variables or window lengths, limiting scalability. Moreover, interpretability and uncertainty estimation remain under‑explored, affecting reliability. To address these gaps, the authors construct a diverse financial image‑text dataset (FVLDB) and develop **Uncertainty‑Adjusted Relative Policy Optimization (UARPO)**, a reinforcement‑learning‑based fine‑tuning method that enables the model to output both predictions and associated uncertainty estimates. **FinZero**, a multimodal pretrained model fine‑tuned with UARPO, performs inference, prediction, and analytical understanding on the FVLDB financial series. Extensive experiments demonstrate strong adaptability and scalability; after UARPO fine‑tuning, FinZero improves prediction accuracy by roughly 13.48 % over GPT‑4o on the high‑confidence subset, confirming the effectiveness of RL‑based fine‑tuning for multimodal large models in financial forecasting.